Senior Research Scientist — Object Pose Estimation & Tracking
Hyunsu Kim
Current focus: Streaming temporal 3D MOT that holds <5cm position error over 4-second occlusions.
Background
PhD KAIST RCV 2020, advised by In So Kweon; thesis on temporally-consistent 6-DoF object pose from monocular driving video. Postdoc at TUM Computer Vision Group with Daniel Cremers (2020–2022), contributing to direct visual odometry baselines. ML Engineer at Motional Perception (2022–2025) shipping the LiDAR-camera tracker for nuScenes-v2 ego routes.
Education
- PhD · Robotics and Computer Vision — KAIST (2020), advisor: In So Kweon
- BS · Mechanical Engineering — POSTECH (2014)
Selected publications
- Kim et al., 'Temporally-Consistent 6-DoF Pose from Monocular Driving Sequences', ICCV 2021
- Kim & Cremers, 'Direct Bundle Adjustment for Multi-Object Pose Tracking', TPAMI 2023
- Kim et al., 'Sparse-to-Dense LiDAR-Camera 3D MOT', CVPR 2024
Reading list
- Xiang et al., 'PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation', RSS 2018
- Yin et al., 'Center-based 3D Object Detection and Tracking (CenterPoint)', CVPR 2021
- Li et al., 'BEVFormer: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers', ECCV 2022
- Weng et al., 'AB3DMOT: A Baseline for 3D Multi-Object Tracking', IROS 2020
- Wang et al., 'StreamPETR: Exploring Object-Centric Temporal Modeling for Multi-View 3D Object Detection', ICCV 2023
Lead experiments
All experiments →- M-001failed
Class-aware Mahalanobis + CTRV in AB3DMOT for AMOTA ≥ 0.55
AMOTA ≥ 0.55 (baseline 0.471)
- M-002done
Differentiable affine + B-spline image registration matching SimpleITK on 5 fixtures (IC-LK + mixed-precision LM + implicit-diff)
rmse_vs_simpleitk_max ≤ 0.0001 (baseline )
- M-005done
Longitudinal mono-modal T1↔T1 registration on real OpenNeuro ds007328 sub-001 (3 sessions, 6-31 days apart)
rmse_vs_simpleitk_max ≤ 0.001 (baseline )